Harvey launches first proprietary legal model Tenet, built on Kimi K3
On August 21, legal AI startup Harvey released Harvey Tenet, its first proprietary legal model post-trained on Moonshot AI's open-weight Kimi K3, alongside six months of research on post-training, evaluation, and open-weight models. The company claims it significantly outperforms general frontier models on legal benchmarks—a notable pivot for a startup that had long relied on Anthropic and OpenAI models.
Confirmed
- Tenet is built on Kimi K3 and fine-tuned with public legal data, synthetic data, and human expert data (m2, m3, m9).
- Per @soumitrashukla9, specialized fine-tuning for high-value economic legal tasks yields substantial gains on Legal Agent Bench, LAB, and LAB Contracts over general frontier models; @lqiao reports Tenet surpasses Opus5 and Fable on multiple dimensions (m3, m4, m5).
- Per @lqiao, Tenet covers 24 corporate law areas and 1,300+ legal tasks (m5).
- New capabilities include long-horizon work such as M&A due diligence (m1, m2).
- Harvey also announced product integration—specialized Agents for litigation, transactions, and in-house work, plus Vault, Knowledge, and a law-firm collaboration ecosystem—with the full legal benchmark report published (m6, m7).
- @pstAsiatech notes Harvey is OpenAI-backed and previously depended on Anthropic/OpenAI models (m9).
Why it matters
- It demonstrates that domain fine-tuning can systematically beat general frontier models on long-horizon, high-value legal tasks, backed by a public benchmark report.
- Post-training on an open-weight base (Kimi K3) with public evaluation offers reproducibility; @markjeffrey relayed views calling this a showcase of open-weight value (m8).
- An OpenAI-backed startup shifting from closed frontier models to an open-weight base is a strong signal for open-weight commercialization.
2026-08-21 ~ 2026-08-22 · 9 related posts
Primary sources
- [source] Harvey unveils Tenet, a model post-trained for long-horizon legal work — rhythmrg · 2026-08-21
- Harvey's Fine-Tuned Model Doubles Performance on Legal Agent Bench, Beating Frontier Models — soumitrashukla9 · 2026-08-21
- Harvey releases Tenet, a post-trained legal model based on Kimi K3 — Exp_Mark · 2026-08-21
- [source] Harvey Launches Tenet Legal Model, Covering 1300+ Tasks — lqiao · 2026-08-21
- Harvey Tenet ships with a full legal-domain benchmark report — lqiao · 2026-08-21
- Harvey Update: AI Agents and Legal Workflow Integration — lqiao · 2026-08-21
- Legal AI startup Harvey launches first model post-trained for law on Kimi K3 — Polymarket · 2026-08-21
- Harvey Launches First Legal AI Model Built on Kimi K3 — markjeffrey · 2026-08-22
- [source] OpenAI-Backed Harvey Builds First Model on Moonshot AI's Kimi K3 — pstAsiatech · 2026-08-22